Cognitive networks plan, decide, and act at different layers of the protocol stack, based on perceived conditions of the network state and assigned rules. We introduce a Cognitive Network Controller (CNC) that relies on a spiking neural network to implement artificial cognition. Based on the outcome of prior actions, the CNC learns how to improve future decisions to maximize its assigned objective. We present the design of the CNC and associated spiking neural network model, a decision encoding strategy based on the time-to- fire of spikes, a learning rule, and synapse weight management. Then, we evaluate the effectiveness of the proposed method using an event-driven simulation of the online selection of end-to-end paths for file transfers over a network of mixed performance. The results indicate that the CNC effectively learns how to dynamically select paths for file transfers that produce lower latency than conventional methods for a broad range of network conditions. The proposed CNC potentially facilitates the integration of cognition and future network applications that require autonomous performance optimization.